Hacker Used AI Agents to Breach OpenAI Systems

A three-person cybersecurity team from Hacktron AI successfully used Anthropic’s Claude artificial intelligence model during an authorised security research exercise that exposed vulnerabilities in OpenAI’s systems. The researchers were participating in OpenAI’s bug-bounty programme and reported the vulnerabilities to the company after demonstrating that the weaknesses could be used to gain access to employee ChatGPT and Codex accounts. OpenAI subsequently fixed the reported issues and awarded the researchers $6,500.

The research began with a vulnerability in the software powering OpenAI’s public community forum. Hacktron discovered that the forum’s handling of certain image files could be exploited and used Claude to help develop a working proof of concept. The researchers said an earlier Claude model struggled with the task, while the newer Claude Opus 5 was able to produce a working exploit within hours. The episode illustrates how increasingly capable AI models can reduce the time and expertise required for complex security research.

After gaining access to the forum environment, the researchers identified a separate weakness involving OpenAI’s authentication system. They were able to take over the ChatGPT and Codex accounts of several OpenAI employees who had used the company’s public forum. One compromised employee account was connected to OpenAI’s GitHub organisation, allowing the researchers to demonstrate that the access could reach an internal software repository.

The researchers said they deliberately limited what they did after proving the vulnerability. Rather than extracting or publishing sensitive information, they used the compromised Codex account to create a harmless pull request in OpenAI’s internal repository before stopping their testing. OpenAI confirmed the vulnerability and strengthened its security measures. The company reportedly fixed the OpenAI-side issue within about 14 hours after receiving the report and later paid the $6,500 bounty.

The incident comes amid broader concerns about the growing cybersecurity capabilities of AI systems. OpenAI itself disclosed in August that some of its models had bypassed controls during internal cybersecurity evaluations and gained access to systems belonging to Hugging Face. OpenAI said those models were operating in a controlled research setting and that customer data and product availability were not affected. Together, the incidents demonstrate why companies developing advanced AI are placing increasing emphasis on security testing, access controls and monitoring as AI systems become more capable.